Pith. sign in

REVIEW 2 major objections 69 references

Interleaved multimodal reasoning works better when models generate only the visual changes between steps, not full images each time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 07:30 UTC pith:OAEQ2Y47

load-bearing objection Clean update-vs-full-image result is real; the 8.4% in-domain headline mixes paradigm with exclusive StructCoT data. the 2 major comments →

arxiv 2607.08434 v1 pith:OAEQ2Y47 submitted 2026-07-09 cs.CV

DeltaV: Thinking with Visual State Updates in Unified Large Multimodal Models

Pengjie Wang , Linger Deng , Zujia Zhang , Shaojie Zhang , Zhenbo Luo , Pei Fu , Jian Luan , Xiang Bai
show 1 more author
Yuliang Liu
This is my paper
classification cs.CV
keywords unified large multimodal modelsinterleaved multimodal reasoningvisual state updatestemporal similarity routertoken allocationStructCoTvisual token efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Unified multimodal models that "think with images" usually regenerate a complete intermediate picture at every reasoning step. This paper argues that is the wrong unit of work: consecutive visual states mostly repeat the same content, so full-image generation wastes tokens and spreads training signal over already-settled pixels instead of the sparse changes that actually move the argument forward. DeltaV instead keeps a base visual state and predicts compact visual-update tokens that capture only what changed. A temporal-similarity router decides how many tokens each update deserves by stopping once further tokens buy little reconstruction gain. The authors also release StructCoT, a 1.05-million-sample dataset spanning 44 domains of structured visual state evolution. On matched experiments the update paradigm cuts newly generated visual tokens by about half while holding reconstruction quality, and it raises multimodal reasoning scores over full-image generation; a 2B-parameter DeltaV trained on StructCoT then beats larger open models on in-domain tests and a same-scale baseline on external benchmarks.

Core claim

The bottleneck in interleaved multimodal reasoning is not pixel-level generation itself but treating every intermediate visual state as independent full-image generation. Modeling reasoning as a base state plus sparse visual updates, with token budgets allocated by temporal similarity and diminishing reconstruction gain, concentrates supervision on reasoning-critical transitions, cuts redundant visual tokens, and improves reasoning.

What carries the argument

Visual state updates with a TSIM Router: each step emits variable-length update tokens conditioned on history; temporal similarity scores group change magnitudes, fitted token–fidelity curves stop allocation when marginal reconstruction gain falls below a threshold, and a learned end token transfers that length policy to inference.

Load-bearing premise

The offline map from how similar successive images look to how many update tokens they need, fitted on one validation set, remains a good general rule for when an update is finished—even at test time when only a learned stop token is available and the true future image is not.

What would settle it

A controlled same-architecture, same-data comparison in which full-image intermediate states match or beat TSIM-routed updates on both reconstruction fidelity and multi-step reasoning accuracy under equal average visual-token budgets would overturn the central efficiency and supervision claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Interleaved multimodal systems can treat visual continuity as a first-class resource instead of paying full-image cost at every step.
  • Training signal for multi-step visual reasoning can be densified on state transitions rather than repeated low-level reconstruction.
  • StructCoT’s scale and seven reasoning-structure categories become a standard training resource for update-centric multimodal models.
  • Adaptive token budgets for visual intermediates become a practical alternative to fixed high-resolution image tokens in unified models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Hybrid pipelines that emit compact updates for small changes and call external tools for deterministic edits (crop, rotate, enhance) follow naturally from the paper’s efficiency argument.
  • The same continuity-plus-delta idea could apply to video, robotics, or GUI agents where successive frames share most pixels.
  • Tasks that need fine local detail may still require higher-capacity perception branches; update tokens alone may under-serve pure recognition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. DeltaV reformulates interleaved multimodal reasoning in ULMMs as progressive visual state updates rather than full-image generation at each step. Conditioned on historical visual tokens, the model predicts compact update tokens ΔZ_t; a TSIM Router (Eqs. 3–9) allocates token budgets from offline-fitted token–SSIM curves by stopping when marginal reconstruction gain falls below τ. The authors also release StructCoT (1.05M samples, 44 domains). Controlled ablations on Zebra-CoT show reconstruction parity at ~64 vs 144 tokens (−55.6% new visual tokens; Fig. 8) and +3.3% reasoning over full-image modeling (Tab. 1). With StructCoT and large-scale data, DeltaV-2B reports +8.4% over larger open models on in-domain tests (Tab. 4) and +5.9% over Qwen3-VL-2B on external benchmarks (Tab. 5).

Significance. If the controlled update-vs-full-image results hold, the paper offers a concrete, efficiency-oriented alternative to the dominant full-image intermediate-state paradigm in ULMMs: denser supervision on sparse transitions, fewer visual tokens, and better temporal consistency (Figs. 10–11). StructCoT is a substantial data contribution for the community. Strengths include matched-architecture ablations (Tabs. 1, 3; Fig. 8), component studies (Tabs. 6–7; Fig. 9), and promised release of code, models, and data. The work is systems-empirical rather than theoretical; its main value is a clear modeling shift plus evidence that reconstruction-aware token routing can improve reasoning under a fixed architecture.

major comments (2)
  1. Abstract and §5.5.1 / Tab. 4 package the +8.4% in-domain gain as if it primarily validates the visual-update paradigm, but the manuscript states that baselines are not fine-tuned on StructCoT. That confounds exclusive access to a 1.05M, 44-domain interleaved corpus with the architectural claim. The load-bearing evidence for the paradigm is the matched Zebra-CoT comparison (+3.3%, Tab. 1; −55.6% tokens, Fig. 8). Please restate headline claims so that data exclusivity is not attributed to visual updates, and add a same-data baseline (e.g., full-image or text-only trained on StructCoT) for Tab. 4-style numbers.
  2. §3.2–3.3 and §5.2: TSIM budgets are calibrated offline on Zebra-CoT validation (DINOv2 features, SSIM curves, fixed τ; Eq. 9) and transferred at inference only via a learned <|vision_end|> token, without online TSIM. Fig. 9(b) and App. A.1 support reconstruction transfer and threshold robustness, but the paper still treats reconstruction fidelity as a proxy for reasoning-useful token allocation (axiom underlying the Router). A short analysis of cases where high reconstruction quality does not improve reasoning (or where the learned stop token under/over-allocates relative to oracle TSIM) would strengthen the central transfer claim.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with offline calibration as design choice, not a derivation that forces claimed gains by construction.

full rationale

DeltaV is an empirical multimodal systems paper. Its load-bearing claims are controlled ablations (visual-update vs full-image under matched architecture/data; Tab. 1, Fig. 8) and external benchmark numbers, not first-principles predictions. The TSIM Router (Sec. 3.2, Eq. 5–9; Sec. 5.2) is an offline calibration that maps temporal similarity intervals to token budgets via fitted SSIM–token curves and a marginal-gain threshold τ; those budgets then supervise training and a learned <|vision_end|> stop token. That is standard hyperparameter/teacher-policy design: reconstruction fidelity at the chosen operating point is intentionally controlled, but the reported reasoning gains (+3.3% over full-image) and external benchmark improvements are measured on held-out tasks whose objectives are not the SSIM curves used for calibration. Fixed-budget visual updates already beat full-image reconstruction and reasoning under the same token count, so the paradigm’s benefit is not forced solely by the slope rule. There is no self-definitional loop (X defined as Y then “derived”), no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known closed-form result. Self-citations, if any, are ordinary related-work context and not load-bearing for the central claims. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

Load-bearing content is mostly engineering assumptions and fitted allocation hyperparameters rather than physical axioms. The central empirical claim rests on visual continuity across reasoning steps, offline reconstruction-based budget selection, and large-scale supervised training on StructCoT plus other multimodal corpora.

free parameters (5)
  • TSIM marginal-gain threshold τ
    Chooses the token budget where dC_m/dx falls below τ (Eq. 9); different τ values produce 36/64/94 average-token operating points (Tab. 6).
  • Temporal decay α in TSIM
    Weights historical states in Eq. 5; ablated around 0.6–0.8 as best (Fig. 9a).
  • Number of TSIM intervals M
    Default M=10 for offline curve fitting; shown relatively robust at 5/10/15 (Tab. 7a).
  • Discrete token budget set K and base K0=144
    Candidate update lengths {9…144} and fixed initial budget define the allocation grid used throughout training and calibration.
  • Loss weights λ_rec, λ_p, λ_adv, λ_sem, λ_vis, β, λ_ent
    Tokenizer and AR training objectives (App. B) balance reconstruction, perceptual, adversarial, VQ, semantic, and visual CE terms.
axioms (4)
  • domain assumption Consecutive intermediate visual states in interleaved reasoning are highly correlated; reasoning-critical information is concentrated in sparse changes.
    Stated in Introduction and Sec. 3 as the motivation for replacing full-image generation with updates.
  • ad hoc to paper Reconstruction fidelity (SSIM/rFID/PSNR under TSIM-Tok) is a valid proxy for allocating tokens that later help reasoning.
    TSIM Router is calibrated purely on reconstruction curves (Sec. 5.2), then transferred as reasoning supervision.
  • domain assumption A frozen foundation visual backbone (SigLIP2) plus discrete update slots can represent task-relevant visual transitions for AR language modeling.
    Core modeling stack in Sec. 3.2 / App. B; standard in modern multimodal tokenizers but still an assumption for reasoning fidelity.
  • ad hoc to paper Offline DINOv2-based temporal similarity is a stable measure of visual change magnitude for budget assignment.
    E is instantiated as DINOv2 only for offline TSIM estimation (Sec. 5.2), separate from the SigLIP2 tokenizer backbone.
invented entities (3)
  • TSIM (temporal similarity) Router no independent evidence
    purpose: Map change magnitude to adaptive visual-update token budgets via offline token–fidelity curves and a slope threshold.
    New allocation mechanism introduced in Sec. 3.2; evidence is internal ablation performance, not an external physical measurement.
  • Visual update tokens ΔZ_t / TSIM-Tok no independent evidence
    purpose: Encode only residual visual changes conditioned on historical states instead of full-image tokens Z_t.
    Central representation of the method; validated by reconstruction and reasoning ablations within the paper.
  • StructCoT dataset no independent evidence
    purpose: Provide 1.05M interleaved multimodal reasoning samples across 44 domains for generalizable update training.
    New resource constructed by the authors; value depends on release quality and community reuse.

pith-pipeline@v1.1.0-grok45 · 37153 in / 3486 out tokens · 35955 ms · 2026-07-10T07:30:31.671061+00:00 · methodology

0 comments
read the original abstract

Current Unified Large Multimodal Models (ULMMs) support interleaved multimodal reasoning through textual reasoning and intermediate visual states, but typically generate each visual state as a full image. This full-image generation paradigm introduces substantial visual-token redundancy and dilutes supervision on sparse yet reasoning-critical state transitions. We propose DeltaV, a ULMM that replaces full-image generation with visual updates. Conditioned on historical visual states, DeltaV incrementally predicts compact update tokens that capture the visual changes across reasoning steps, avoiding repeated modeling of unchanged content. To align the token budget of each update with the magnitude of visual change, DeltaV introduces a temporal similarity (TSIM) Router, which stops allocating tokens once the marginal reconstruction gain falls below a threshold. To support more diverse and generalizable reasoning, we further construct StructCoT, a large-scale interleaved multimodal reasoning dataset with 1.05M samples spanning 44 task domains. Experiments show that the visual-update paradigm reduces newly generated visual tokens by 55.6\% on average without compromising reconstruction fidelity, and improves multimodal reasoning by 3.3\% over full-image generation. Trained with StructCoT and large-scale multimodal data, DeltaV-2B further outperforms substantially larger open-source models by 8.4\% on in-domain multimodal reasoning evaluations and surpasses the comparable-scale Qwen3-VL-2B by 5.9\% on external multimodal reasoning and understanding benchmarks. Code, models, and StructCoT will be released at https://github.com/Pengjie-W/DeltaV.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · 13 internal anchors

  1. [1]

    Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others , journal=

    work page

  2. [2]

    Wu, Zhiyu and Chen, Xiaokang and Pan, Zizheng and Liu, Xingchao and Liu, Wen and Dai, Damai and Gao, Huazuo and Ma, Yiyang and Wu, Chengyue and Wang, Bingxuan and others , journal=

    work page

  3. [3]

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and others , journal=

    work page

  4. [4]

    Yunzhuo Hao and Jiawei Gu and Huichen Will Wang and Linjie Li and Zhengyuan Yang and Lijuan Wang and Yu Cheng , booktitle=. Can

    work page

  5. [5]

    Dongzhi Jiang and Renrui Zhang and Ziyu Guo and Yanwei Li and Yu Qi and Xinyan Chen and Liuhui Wang and Jianhan Jin and Claire Guo and Shen Yan and Bo Zhang and Chaoyou Fu and Peng Gao and Hongsheng Li , booktitle=

    work page

  6. [6]

    Image-of-Thought Prompting for Visual Reasoning Refinement in Multimodal Large Language Models

    Image-of-thought prompting for visual reasoning refinement in multimodal large language models , author=. arXiv preprint arXiv:2405.13872 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Zheng, Ziwei and Yang, Michael and Hong, Jack and Zhao, Chenxiao and Xu, Guohai and Yang, Le and Shen, Chao and Yu, Xing , journal=

    work page

  8. [8]

    Hu, Yushi and Shi, Weijia and Fu, Xingyu and Roth, Dan and Ostendorf, Mari and Zettlemoyer, Luke and Smith, Noah A and Krishna, Ranjay , booktitle=

    work page

  9. [9]

    Chen, Xiaokang and Wu, Zhiyu and Liu, Xingchao and Pan, Zizheng and Liu, Wen and Xie, Zhenda and Yu, Xingkai and Ruan, Chong , journal=

    work page

  10. [10]

    Emerging Properties in Unified Multimodal Pretraining

    Emerging properties in unified multimodal pretraining , author=. arXiv preprint arXiv:2505.14683 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Li, Ang and Wang, Charles and Fu, Deqing and Yue, Kaiyu and Cai, Zikui and Zhu, Wang Bill and Liu, Ollie and Guo, Peng and Neiswanger, Willie and Huang, Furong and others , journal=

    work page

  12. [12]

    Gu, Jiawei and Hao, Yunzhuo and Wang, Huichen Will and Li, Linjie and Shieh, Michael Qizhe and Choi, Yejin and Krishna, Ranjay and Cheng, Yu , journal=

    work page

  13. [13]

    Proceedings of the International Conference on Machine Learning , year=

    Imagine While Reasoning in Space: Multimodal Visualization-of-Thought , author=. Proceedings of the International Conference on Machine Learning , year=

    work page

  14. [14]

    Advances in Neural Information Processing Systems , year=

    Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation , author=. Advances in Neural Information Processing Systems , year=

    work page

  15. [15]

    Du, Sinan and Guo, Jiahao and Li, Bo and Cui, Shuhao and Xu, Zhengzhuo and Luo, Yifu and Wei, Yongxian and Gai, Kun and Wang, Xinggang and Wu, Kai and others , journal=

    work page

  16. [16]

    He, Xin and Wei, Longhui and Ouyang, Jianbo and Xie, Lingxi and Tian, Qi , journal=

    work page

  17. [17]

    Emu3.5: Native Multimodal Models are World Learners

    Emu3.5: Native multimodal models are world learners , author=. arXiv preprint arXiv:2510.26583 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    and Krishna, Ranjay , booktitle=

    Bigverdi, Mahtab and Luo, Zelun and Hsieh, Cheng-Yu and Shen, Ethan and Chen, Dongping and Shapiro, Linda G. and Krishna, Ranjay , booktitle=. Perception Tokens Enhance Visual Reasoning in Multimodal Language Models , year=

    work page

  19. [19]

    Introducing Visual Perception Token into Multimodal Large Language Model

    Introducing visual perception token into multimodal large language model , author=. arXiv preprint arXiv:2502.17425 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Visual programming: Compositional visual reasoning without training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  21. [21]

    Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning

    Pixel reasoner: Incentivizing pixel-space reasoning with curiosity-driven reinforcement learning , author=. arXiv preprint arXiv:2505.15966 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Interleaved-modal chain-of-thought , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  23. [23]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Whiteboard-of-thought: Thinking step-by-step across modalities , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2024

  24. [24]

    Reinforced Visual Perception with Tools

    Reinforced visual perception with tools , author=. arXiv preprint arXiv:2509.01656 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    2025 , month =

    Thinking with images , howpublished =. 2025 , month =

    work page 2025

  26. [26]

    Chen, Zhangquan and Zhao, Ruihui and Luo, Chuwei and Sun, Mingze and Yu, Xinlei and Kang, Yangyang and Huang, Ruqi , booktitle=

    work page

  27. [27]

    Zhao, Qingqing and Lu, Yao and Kim, Moo Jin and Fu, Zipeng and Zhang, Zhuoyang and Wu, Yecheng and Li, Zhaoshuo and Ma, Qianli and Han, Song and Finn, Chelsea and others , booktitle=

    work page

  28. [28]

    Shi, Weikang and Yu, Aldrich and Fang, Rongyao and Ren, Houxing and Wang, Ke and Zhou, Aojun and Tian, Changyao and Fu, Xinyu and Hu, Yuxuan and Lu, Zimu and others , journal=

    work page

  29. [29]

    Tong, Shengbang and Fan, David and Li, Jiachen and Xiong, Yunyang and Chen, Xinlei and Sinha, Koustuv and Rabbat, Michael and LeCun, Yann and Xie, Saining and Liu, Zhuang , booktitle=

    work page

  30. [30]

    Lin, Bin and Li, Zongjian and Cheng, Xinhua and Niu, Yuwei and Ye, Yang and He, Xianyi and Yuan, Shenghai and Yu, Wangbo and Wang, Shaodong and Ge, Yunyang and others , journal=

    work page

  31. [31]

    Wu, Chenyuan and Zheng, Pengfei and Yan, Ruiran and Xiao, Shitao and Luo, Xin and Wang, Yueze and Li, Wanli and Jiang, Xiyan and Liu, Yexin and Zhou, Junjie and others , journal=

    work page

  32. [32]

    Fang, Rongyao and Duan, Chengqi and Wang, Kun and Li, Hao and Huang, Linjiang and Tian, Hao and Zeng, Xingyu and Zhao, Rui and Dai, Jifeng and Li, Hongsheng and others , booktitle=

    work page

  33. [33]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  34. [34]

    Mogao: An Omni Foundation Model for Interleaved Multi-Modal Generation

    Mogao: An omni foundation model for interleaved multi-modal generation , author=. arXiv preprint arXiv:2505.05472 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    Yang, Ling and Tian, Ye and Li, Bowen and Zhang, Xinchen and Shen, Ke and Tong, Yunhai and Wang, Mengdi , booktitle=

    work page

  36. [36]

    Li, Shufan and Gu, Jiuxiang and Liu, Kangning and Lin, Zhe and Wei, Zijun and Grover, Aditya and Kuen, Jason , journal=

    work page

  37. [37]

    Xin, Yi and Qin, Qi and Luo, Siqi and Zhu, Kaiwen and Yan, Juncheng and Tai, Yan and Lei, Jiayi and Cao, Yuewen and Wang, Keqi and Wang, Yibin and others , journal=. Lumina-

    work page

  38. [38]

    arXiv preprint arXiv:2511.23469 , year=

    Visual Generation Tuning , author=. arXiv preprint arXiv:2511.23469 , year=

    work page arXiv

  39. [39]

    2024 , organization=

    Liu, Yuan and Duan, Haodong and Zhang, Yuanhan and Li, Bo and Zhang, Songyang and Zhao, Wangbo and Yuan, Yike and Wang, Jiaqi and He, Conghui and Liu, Ziwei and others , booktitle=. 2024 , organization=

    work page 2024

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Are we on the right way for evaluating large vision-language models? , author=. Advances in Neural Information Processing Systems , volume=

    work page

  41. [41]

    Zou, Kai and Huang, Ziqi and Dong, Yuhao and Tian, Shulin and Zheng, Dian and Liu, Hongbo and He, Jingwen and Liu, Bin and Qiao, Yu and Liu, Ziwei , journal=

    work page

  42. [42]

    Beyond language modeling: An exploration of multimodal pretraining, March 2026

    Beyond language modeling: An exploration of multimodal pretraining , author=. arXiv preprint arXiv:2603.03276 , year=

    work page arXiv

  43. [43]

    Zhang, Haichao and Li, Yijiang and He, Shwai and Nagarajan, Tushar and Chen, Mingfei and Lu, Jianglin and Li, Ang and Fu, Yun , journal=

    work page

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation , author=. Advances in Neural Information Processing Systems , volume=

    work page

  45. [45]

    Qwen2.5-VL Technical Report

    Qwen2.5-VL Technical Report , author=. arXiv preprint arXiv:2502.13923 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    Chern, Ethan and Su, Jiadi and Ma, Yan and Liu, Pengfei , journal=

    work page

  47. [47]

    Chameleon: Mixed-Modal Early-Fusion Foundation Models

    Chameleon: Mixed-modal early-fusion foundation models , author=. arXiv preprint arXiv:2405.09818 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Hurst, Aaron and Lerer, Adam and Goucher, Adam P and Perelman, Adam and Ramesh, Aditya and Clark, Aidan and Ostrow, AJ and Welihinda, Akila and Hayes, Alan and Radford, Alec and others , journal=

    work page

  49. [49]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  50. [50]

    arXiv preprint arXiv:2505.02567 , year=

    Unified multimodal understanding and generation models: Advances, challenges, and opportunities , author=. arXiv preprint arXiv:2505.02567 , year=

    work page arXiv

  51. [51]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, Maxime and Darcet, Timoth. arXiv preprint arXiv:2304.07193 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  52. [52]

    Singh, Aaditya and Fry, Adam and Perelman, Adam and Tart, Adam and Ganesh, Adi and El-Kishky, Ahmed and McLaughlin, Aidan and Low, Aiden and Ostrow, AJ and Ananthram, Akhila and others , journal=

    work page

  53. [53]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Monkey: Image resolution and text label are important things for large multi-modal models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  54. [54]

    arXiv preprint arXiv:2602.08524 , year=

    GeoFocus: Blending Efficient Global-to-Local Perception for Multimodal Geometry Problem-Solving , author=. arXiv preprint arXiv:2602.08524 , year=

    work page arXiv

  55. [55]

    International Journal of Computer Vision , volume=

    Liquid: Language models are scalable and unified multi-modal generators , author=. International Journal of Computer Vision , volume=. 2026 , publisher=

    work page 2026

  56. [56]

    Synthetic Similarity Search in Automotive Production

    Synthetic Similarity Search in Automotive Production , author=. arXiv preprint arXiv:2505.07256 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence , author=. Advances in Neural Information Processing Systems , volume=

    work page

  58. [58]

    Vision Foundation Models as Generalist Tokenizers for Image Generation

    Vision Foundation Models as Generalist Tokenizers for Image Generation , author=. arXiv preprint arXiv:2605.18390 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    Du, Sinan and Guo, Jiahao and Li, Bo and Cui, Shuhao and Xu, Zhengzhuo and Luo, Yifu and Wei, Yongxian and Gai, Kun and Wang, Xinggang and Wu, Kai and others , booktitle=

    work page

  60. [60]

    Tschannen, Michael and Gritsenko, Alexey and Wang, Xiao and Naeem, Muhammad Ferjad and Alabdulmohsin, Ibrahim and Parthasarathy, Nikhil and Evans, Talfan and Beyer, Lucas and Xia, Ye and Mustafa, Basil and others , journal=

    work page

  61. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Machine mental imagery: Empower multimodal reasoning with latent visual tokens , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  62. [62]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Monet: Reasoning in Latent Visual Space Beyond Image and Language , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  63. [63]

    Update to GPT-5 System Card: GPT-5.2 , year =

    work page

  64. [64]

    Chen, Qiguang and Qin, Libo and Zhang, Jin and Chen, Zhi and Xu, Xiao and Che, Wanxiang , booktitle=

    work page

  65. [65]

    Lu, Pan and Bansal, Hritik and Xia, Tony and Liu, Jiacheng and Li, Chunyuan and Hajishirzi, Hannaneh and Cheng, Hao and Chang, Kai-Wei and Galley, Michel and Gao, Jianfeng , booktitle=

    work page

  66. [66]

    Xu, Weiye and Wang, Jiahao and Wang, Weiyun and Chen, Zhe and Zhou, Wengang and Yang, Aijun and Lu, Lewei and Li, Houqiang and Wang, Xiaohua and Zhu, Xizhou and Wang, Wenhai and Dai, Jifeng and Zhu, Jinguo , booktitle=

    work page

  67. [67]

    Fu, Chaoyou and Chen, Peixian and Shen, Yunhang and Qin, Yulei and Zhang, Mengdan and Lin, Xu and Yang, Jinrui and Zheng, Xiawu and Li, Ke and Sun, Xing and Wu, Yunsheng and Ji, Rongrong and Shan, Caifeng and He, Ran , booktitle=

    work page

  68. [68]

    Wu, Penghao and Xie, Saining , booktitle=

    work page

  69. [69]

    Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal

    Tong, Shengbang and Liu, Zhuang and Zhai, Yuexiang and Ma, Yi and LeCun, Yann and Xie, Saining , booktitle=. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal

    work page